在工具调用过程中，如何将中间状态返回到的stream，可以使用from langchain_core.callbacks import dispatch_custom_event方法实现。示例如下

from langchain_core.callbacks import dispatch_custom_event
from langchain_core.tools import tool

@tool
def query_and_summary(question: str):
    """
    根据用户提问，查询相关文档，并对命中的文档片段进行总结
    """
    # 模拟结构化数据
    docs = {"《听懂掌声》": "听懂掌声", "《成功学》": "自信&成功"}
    dispatch_custom_event(name="docs", data=docs)
    # 模拟流式数据
    for data in summary():
        dispatch_custom_event(name="summary", data=data)

def summary():
    summary = "“听懂掌声”这个词语是来自于一个培训视频，在视频里面做培训的人叫做枭哥"
    for data in summary:
        yield data

• graph定义

这里我们定义了一个模拟查询文档并返回总结的工具，graph调用时需通过astream_events方法调用。当执行dispatch_custom_event函数时，stream会收到一个on_custom_event事件，事件的name是我们在函数中传入的name

from typing_extensions import Annotated, TypedDict

from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables import Runnable, RunnableConfig
from langchain_core.tools import tool
from langgraph.graph.message import AnyMessage, add_messages
from langgraph.graph import StateGraph, START, END
from langgraph.prebuilt import ToolNode, InjectedState, tools_condition
from langgraph.checkpoint.memory import MemorySaver

class State(TypedDict):
    messages: Annotated[list[AnyMessage], add_messages]

class Assistant:
    def __init__(self, runnable: Runnable):
        self.runnable = runnable

    def __call__(self, state: State, config: RunnableConfig):
        result = self.runnable.invoke(state)
        return {"messages": result}

assistant_system_prompt = "你是成功学助手，收到用户提问请调用query_and_summary工具"
assistant_prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            assistant_system_prompt,
        ),
        ("placeholder", "{messages}"),
    ],
)
assistant_tool = [query_and_summary]
assistant_runnable = assistant_prompt | llm.bind_tools(assistant_tool)
assistant = Assistant(assistant_runnable)

# node定义
builder = StateGraph(State)
builder.add_node("assistant", assistant)
builder.add_node("tools", ToolNode(assistant_tool))

# edge定义
builder.add_edge(START, "assistant")
builder.add_conditional_edges("assistant", tools_condition)
builder.add_edge("tools", END)
memory = MemorySaver()
graph = builder.compile(checkpointer=memory)

• 调用

import uuid
thread_id = str(uuid.uuid4())
config = {"configurable": {"thread_id": thread_id}}
async for chunk in graph.astream_events({"messages": ("user", "什么是成功学")}, config=config, version="v2"):
    if chunk["event"] == "on_custom_event" and chunk["name"] == "docs":
        print(chunk["data"])
    if chunk["event"] == "on_custom_event" and chunk["name"] == "summary":
        print("recv:" + chunk["data"])

{'《听懂掌声》': '听懂掌声', '《成功学》': '自信&成功'}
recv:“
recv:听
recv:懂
recv:掌
recv:声
recv:”
recv:这
recv:个
recv:词
recv:语
recv:是
...